Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Composer: Building a fast frontier model with RL

cursor.com

Post reply on HN
Print view View original

111–120 of 169 posts

Re: Composer: Building a fast frontier model with RL

#111

Post by srush »

smg wrote: post #106
srush wrote: post #16

Hi everyone, I am an ML researcher at Cursor, and worked on this project. Would love to hear any feedback you may have on the model, and can answer question about the blog post.

Can you please tell us more about how you used Ray for setting up the RL infrastructure?

Oh good question. Actually speaking at the Ray Summit next week in SF so we will talk more about it. We used Ray throughout the pipeline for running evals, for the RL controller, for data collation, and for visualizations. One tool we found helpful was Ray Data which let us easily scale over data and run logs.

Re: Composer: Building a fast frontier model with RL

#112

Post by nvartolomei »

srush wrote: post #111
smg wrote: post #106

Earlier quoted context omitted.

Can you please tell us more about how you used Ray for setting up the RL infrastructure?

Oh good question. Actually speaking at the Ray Summit next week in SF so we will talk more about it. We used Ray throughout the pipeline for running evals, for the RL controller, for data collation, and for visualizations. One tool we found helpful was Ray Data which let us easily scale over data and run logs.

Please share more about Ray Data use case.

Re: Composer: Building a fast frontier model with RL

#113

Post by embedding-shape »

srush wrote: post #16

Hi everyone, I am an ML researcher at Cursor, and worked on this project. Would love to hear any feedback you may have on the model, and can answer question about the blog post.

Do you have any graphs handy that kind of replicates the one used first in the blog post but a bit less ambiguous, maybe without model grouping? I feel like it would have been a bit more fair to include proper names, and individualize them rather than group everything together by something, and then present your own model on its own.

Re: Composer: Building a fast frontier model with RL

#114

Post by Jcampuzano2 »

netcraft wrote: post #107

I love cursor, the tab completion and agent mode. But I really dislike vscode after using intellij for so many years. I really wish the underlying editor was better, or I could get cursor features in intellij instead. The editing of the files is mostly fine, but its everything else around it that a full IDE provides thats just so much better. Right now its intellij + claude code for me, and its fine, but I wish I cou…

Building off of VSCode was probably Cursors silver bullet and the best decision they could have ever made.

It made migrating for everyone using VSCode (probably the single most popular editor) or another vscode forked editor (but at the time it was basically all VSCode) as simple as install and import settings.

I do not think Cursor would have done nearly as well as it has if it didn't. So even though it can be subpar in some areas due to VSCodes baggage, its probably staying that way for a while.

Re: Composer: Building a fast frontier model with RL

#116

Post by srush »

nvartolomei wrote: post #112
srush wrote: post #111

Earlier quoted context omitted.

Oh good question. Actually speaking at the Ray Summit next week in SF so we will talk more about it. We used Ray throughout the pipeline for running evals, for the RL controller, for data collation, and for visualizations. One tool we found helpful was Ray Data which let us easily scale over data and run logs.

Please share more about Ray Data use case.

We use Ray data for our map-style processing jobs. For example one tool have runs over all the rollouts from the RL system and collects qualitative statistics to understand which type of agent trajectories are being reward, and what types of searches and terminal commands are being made.

Re: Composer: Building a fast frontier model with RL

#117

Post by netcraft »

Jcampuzano2 wrote: post #114
netcraft wrote: post #107

I love cursor, the tab completion and agent mode. But I really dislike vscode after using intellij for so many years. I really wish the underlying editor was better, or I could get cursor features in intellij instead. The editing of the files is mostly fine, but its everything else around it that a full IDE provides thats just so much better. Right now its intellij + claude code for me, and its fine, but I wish I cou…

Building off of VSCode was probably Cursors silver bullet and the best decision they could have ever made. It made migrating for everyone using VSCode (probably the single most popular editor) or another vscode forked editor (but at the time it was basically all VSCode) as simple as install and import settings. I do not think Cursor would have done nearly as well as it has if it didn't. So even though it can be subpa…

I dont disagree with anything you said. If I was in their shoes, I would have done exactly the same thing.

Maybe my complaint is that I wish vscode had more features like intellij, or that intellij was the open source baseline a lot of other things could be built on.

Intellij is not without its cruft and problems, dont get me wrong. But its git integration, search, navigation, database tools - I could go on - all of these features are just so much nicer than what vscode offers.

Re: Composer: Building a fast frontier model with RL

#118

Post by srush »

timcobb wrote: post #102
matheist wrote: post #4

Earlier quoted context omitted.

> "Best Frontier" includes GPT-5 and Sonnet 4.5, which both outperform Composer.

Does anyone code with GPT-5? I've never had it work in Cursor. I mean, like, at all.

A lot of people use it! It scores very well on our benchmarks, significantly better than Composer-1.

Re: Composer: Building a fast frontier model with RL

#119

Post by srush »

gabriel666smith wrote: post #101
srush wrote: post #76

Earlier quoted context omitted.

Unfortunately not, as we used our own internal code for the benchmark. We would also like to see more benchmarks that reflect the day-to-day agentic coding use.

Is there any information at all available, anywhere, on what Cursor Bench is testing and how? It's the most prominent part of the release post - but it's really hard to understand what exactly it's saying.

Roughly, we had Cursor software engineers record real questions they were asking models, and then had them record the PR that they made that contained the result. We then cleaned these up. That is the benchmark.

Re: Composer: Building a fast frontier model with RL

#120

Post by ukblewis »

srush wrote: post #119
gabriel666smith wrote: post #101

Earlier quoted context omitted.

Is there any information at all available, anywhere, on what Cursor Bench is testing and how? It's the most prominent part of the release post - but it's really hard to understand what exactly it's saying.

Roughly, we had Cursor software engineers record real questions they were asking models, and then had them record the PR that they made that contained the result. We then cleaned these up. That is the benchmark.

Which programming languages/tools/libraries did the teams questions/code involve?
Post reply on HN